A Hybrid AI Catches More Rare Diseases Without Dropping the Receipts

Patients with rare diseases often wait years and see more than a dozen specialists before a correct diagnosis is named. A new arXiv preprint proposes a hybrid system that lifts the first-pass accuracy of an established diagnostic tool by 5 to 20 percentage points and keeps the structured evidence that a clinician can audit.

The system pairs a large language model (LLM) with an ontology ranker, a curated database that maps patient symptoms to diseases and shows its work. The LLM produces a broad, flexible differential diagnosis. The ontology ranker produces a narrower, evidence-anchored one. The new contribution is a fusion model that examines the two ranked lists, the agreement between them, and the ontology's per-candidate support, and learns, per case, how much to lean on the LLM.

Across eight open LLMs, the fusion lifts the first-guess accuracy (Recall@1) of the ontology ranker Phenomizer by 7.86 percentage points on the Phenopacket Store benchmark and by 20.18 points on the smaller RAMEDIS benchmark, per the paper. The preprint was flagged by the TLDR aggregator.

The two benchmarks differ in scale and difficulty. Phenopacket Store is the larger, more curated set; RAMEDIS is smaller and harder, where standalone Phenomizer often misses the right disease on the first guess. The 7.86 point lift on Phenopacket Store is a meaningful gain on a benchmark where top-1 is already non-trivial. The 20.18 point lift on RAMEDIS is the larger swing, and the more relevant number for a clinician thinking about cases where the standard tool is currently weak.

In a separate test using a single LLM, DeepSeek-V4-Flash, served via API and never retrained for the task, Recall@1 moved from 0.1657 to 0.2176, a 5.19-point gain. That test matters because the fusion model was trained only on the other LLMs. The 5.19 point lift on an unseen API-served model is the closest thing in the paper to a generalization result.

The frame the authors choose matters. It is non-replacement: the LLM strengthens the ranker rather than substituting for it. In practice, that means the structured evidence from the ontology stays attached to most of the top-1 answers. For 90.8% of correct fused diagnoses, the disease retains candidate-level ontology evidence that a clinician can inspect. That number is the trust trailer of the paper: a hybrid that wins on accuracy and still shows the receipts.

The authors had to remove a contamination pathway in the benchmark. Phenopacket Store cases and the ontology annotations used to score them were derived from the same publications, and that overlap would have inflated apparent gains. The team identified the leak, fixed it, and reran. The numbers reported above are post-fix.

The mechanism is also a small counter to the dominant 2026 frame, which treats clinical AI as a contest between rules-based knowledge systems and end-to-end LLMs. This work argues for a third lane: learn, per case, how to weight the two. The authors describe it as a behavior-based fusion because the fusion model watches how each system behaves on a case and adapts its reliance accordingly.

There is no prospective validation, no workflow integration, and no patient-outcome data. The 5 to 20 point gain is a benchmark gain. A hybrid can outperform either system alone without sacrificing the evidence trail. For a field under pressure to explain its decisions, that is a usable shape.

Next milestone to watch: whether the fusion approach generalizes across other API-served LLMs and whether a peer-reviewed version of the paper carries the test-set leakage fix into its official benchmarks.